Skip to content

CAM: A Constructivist View of Agentic Memory for LLM-Based Reading Comprehension

🕒 Published (v1): 2025-01-01 · Source: NeurIPS · Venue: NeurIPS 2025 · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

CAM (Constructivist Agentic Memory) applies Jean Piaget's Constructivist Theory to design a principled agentic memory system for LLM-based long-document reading comprehension. It combines hierarchical memory structure with an incremental overlapping clustering algorithm that supports both flexible assimilation (many-to-many chunk-to-abstraction mapping) and dynamic accommodation (local-only updates). CAM outperforms all baselines on six benchmarks while being 4× faster than offline methods in online batch settings.

Problem

Existing LLM memory modules for long-text comprehension lack a systematic design principle. Prior work either uses unstructured tabular memory (MemGPT, ReadAgent) that misses cross-chunk dependencies, or structured memory systems (RAPTOR, GraphRAG, MemTree) that fail to simultaneously satisfy flexibility (many-to-many chunk affiliation) and dynamicity (incremental updates without full reconstruction). No existing method satisfies all three desiderata: structurality, flexibility, and dynamicity.

Method

CAM grounds its design in Piaget's three cognitive traits—structured schemata, flexible assimilation, dynamic accommodation—and implements them via a three-step incremental overlapping clustering algorithm:

  1. Foundational Network Expansion: New text chunks are integrated into a semantic graph G₀ using a composite similarity score combining cosine embedding similarity and Gaussian positional proximity (weighted by α). Each chunk connects to top-k neighbors exceeding a threshold θ.

  2. Ego-Centric Disentanglement: For each affected node, CAM extracts its ego-network (neighbors minus self), partitions it into connected components, and replicates the node once per component. This converts the overlapping structure into a non-overlapping replica graph G̃₀, enabling flexible many-to-many affiliation to higher-level abstractions.

  3. Online Clustering Updates: An incremental label propagation algorithm runs on G̃₀, updating only affected nodes. Converged clusters trigger LLM-based summarization to produce abstraction supernodes at the next layer, which then recursively trigger steps 2–3 upward.

At inference, a Prune-and-Grow strategy first globally retrieves top-s candidate nodes by embedding similarity, then uses an LLM to iteratively activate neighbors and children of relevant nodes until the activation set stabilizes. This combines global semantic matching with local structural traversal.

Key Contributions

  • Explicit design blueprint for agentic memory grounded in Piagetian constructivism: three formalized traits (structured schemata, flexible assimilation, dynamic accommodation)
  • CAM prototype implementing all three traits simultaneously via incremental overlapping clustering—the first online system to do so
  • Prune-and-Grow associative retrieval that outperforms both hierarchical traversal and global retrieval strategies
  • Batch-level online integration with sublinear time scaling and parallelizable replication

Results

All results from Table 2 (GPT-4o-mini backbone, text-embedding-3-small):

  • NovelQA (QA): CAM R-L 25.4 vs. best baseline RAPTOR 23.7; ACC-L 52.3 vs. RAPTOR 47.8
  • QMSum (summarization): CAM R-L 26.5 vs. GraphRAG 25.2; ACC-L 57.6 vs. GraphRAG 53.9
  • FABLES (claim verification): CAM F1P 91.5 vs. GraphRAG 87.2; F1N 52.5 vs. RAPTOR 48.5
  • MultiHop-RAG (multi-doc QA): CAM EM 72.8, F1 77.5 vs. RAPTOR 69.4/73.6
  • ODSum-Story: CAM ACC-L 54.6 vs. GraphRAG 50.2
  • ODSum-Meeting: CAM ACC-L 50.7 vs. GraphRAG 45.8
  • Average gain of +3.0% across all metrics over best baselines
  • Online batch insertion: CAM <0.3h regardless of batch size vs. RAPTOR/GraphRAG >1.2h; MemTree exceeds offline reconstruction at batch size >400 chunks
  • Online CAM performance remains stable across batch sizes 1–500, remaining competitive with offline baselines

Limitations

  • Scope restricted to reading comprehension tasks; behavioral planning, long-sequence generation, and multimodal settings unexplored
  • No self-questioning or reflection behaviors; agentic behaviors beyond memory are absent
  • Hallucinations in LLM-generated summaries propagate upward through the hierarchy and are not detected or mitigated
  • Assumes internally consistent source documents; contradictory or conflicting information across sources is not handled
  • Relies on fixed prompts and tuned hyperparameters; no trainable memory controller or adaptive update policy
  • Current implementation (incremental overlapping clustering) is one instantiation of the constructivist principle; alternatives (neural controllers, symbolic planners) not explored

Relevance to Agentic AI / LLM Agents

CAM directly addresses a core infrastructure challenge for LLM agents: how to maintain a coherent, updatable long-term memory over documents that exceed context windows. By grounding memory design in cognitive theory rather than ad hoc heuristics, it provides a principled framework applicable beyond RAG—relevant to any agent that must accumulate and query episodic or knowledge memory over streaming inputs. The batch-online capability (4× speedup over offline reconstruction) is particularly significant for agents operating in real-time or serial reading scenarios (e.g., tracking evolving documents, news feeds). The Prune-and-Grow retrieval mechanism also resembles associative recall patterns studied in cognitive architectures, offering a concrete bridge between cognitive science and agent memory design that could inform future work on planning and multi-modal agents.